Exposure outcome confounder effect modifier

Exposure, outcome, confounder, effect modifier
● Differential vs. non-differential Misclassification of exposure: “Differential” is with respect to the outcome
Misclassification of outcome: “Differential” is with respect to exposure

Misclassification of exposure
Non-differential
● You have incorrectly measured exposure, but the extent to which you have done this is the
same in those with and without the outcome
● Example: case-control study, exposure is measured via a biomarker, and there is systematic
error in your assay. We don’t expect this to be different in cases versus controls
● Effect: bias towards the null
Differential:
● Exposure ascertainment results in a different degree of measurement error in those with the
outcome versus those without the outcome
● Example: case-control study, exposure is measured via a biomarker, and the lab has an
extra-qualified technician process the case samples
● Effect: can’t predict, can go all over the place

Misclassification of outcome
Non-differential
● You have incorrectly measured outcome, but the extent to which you have done this is the
same in those with and without the exposure
● Example: case-control study, and it’s impossible to screen all of the controls to make sure
none of them are a case (e.g. this would require invasive surgery)
● Effect: bias towards the null*
Differential:
● Outcome ascertainment results in a different degree of measurement error in the exposed
versus the unexposed
● Example: cohort study, exposed receive extra scrutiny
● Effect: can’t predict, can go all over the place
*In a follow-up study, measurement of outcome with imperfect sensitivity actually doesn’t produce bias, but it does compromise power

Misclassification of confounder
Always results in residual confounding!
● Direction of bias will be the same as the direction of bias due to confounding
● Your adjusted will be between the “crude” and the “truth”
We don’t tend to talk about this as differential versus non-differential

Misclassification
Things to “get straight”
● What are you misclassifying?
● Differential vs. non-differential
● Connections to sensitivity and specificity
sensitivity <100%: we have falsely classified some
people as not having that outcome, exposure, or
confounder value, who actually do
specificity <100%: we have falsely classified some
people as having that value when they do not

Why does non-differential misclassification bias
towards the null?
Sensitivity <100%, exposure= obesity, outcome= heart attack:
● We’re putting a bunch of obese people in the non-obese group
● If there is an association between obesity and heart attack, we’ll be making the
non-obese people look like they have a higher risk of heart attack than they really do
Specificity <100%, exposure= obesity, outcome= heart attack:
● We’re putting a bunch of non-obese people in the obese group
● If there is an association, we’ll be making the obese people look like they have a lower
risk of heart attack than they really do